Is this you? Claim this profile to correct it, add a bio and choose the work people see first.
Claim this profileWorks133 from public data
- Perceptual-Sensitive GAN for Generating Adversarial Patches369
This paper proposes a perceptual-sensitive generative adversarial network (PS-GAN) that can simultaneously enhance the visual fidelity and the attacking ability for the adversarial patch, and treats the patch generation as a patch-to-patch translation via an adversarial process.
- Dual Attention Suppression Attack: Generate Adversarial Camouflage in Physical World282
The Dual Attention Suppression (DAS) attack is proposed to generate visually-natural physical adversarial camouflages with strong transferability by suppressing both model and human attention and outperforms state-of-the-art methods.
- Bias-Based Universal Adversarial Patch Attack for Automatic Check-Out170
A bias-based framework to generate class-agnostic universal adversarial patches with strong generalization ability, which exploits both the perceptual and semantic bias of models.
- BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning145
The BadCLIP attack is introduced, which is resistant to backdoor detection and model fine-tuning defenses and draws motivations from the perspective of the Bayesian rule and proposes a dual-embedding guided framework for backdoor attacks.
- Training Robust Deep Neural Networks via Adversarial Noise Propagation145
A simple yet powerful training algorithm, named Adversarial Noise Propagation (ANP), which injects noise into the hidden layers in a layer-wise manner and enables the strong robustness for deep models against both adversarial and corrupted ones, and also significantly outperforms various adversarial defense methods.
- Jailbreak Vision Language Models via Bi-Modal Adversarial Prompt140
The Bi-Modal Adversarial Prompt Attack (BAP) is introduced, which executes jailbreaks by optimizing textual and visual prompts cohesively and significantly outperforms other methods by large margins.
- 125
- BiBERT: Accurate Fully Binarized BERT124
BiBERT, an accurate fully binarized BERT, is proposed to eliminate the performance bottlenecks and introduces an efficient Bi-Attention structure for maximizing representation information statistically and a Direction-Matching Distillation scheme to optimize the full binarization BERT accurately.
- A comprehensive evaluation framework for deep model robustness105
This paper establishes a model robustness evaluation framework containing 23 comprehensive and rigorous metrics, which consider two key perspectives of adversarial learning (i.e., data and model) and provides a comprehensive evaluation framework, where researchers could conduct comprehensive and fast evaluations using the open-source toolkit.
- X-Adv: Physical Adversarial Object Attacks against X-ray Prohibited Item Detection102
This paper proposes X-adv to generate physically printable metals that act as an adversarial agent capable of deceiving X-ray detectors when placed in luggage, and develops a differentiable converter that facilitates the generation of 3D-printable objects with adversarial shapes.
- Spatiotemporal Attacks for Embodied Agents99
This work generates spatiotemporal perturbations to form 3D adversarial examples, which exploit the interaction history in both the temporal and spatial dimensions and adversarially perturb the physical properties of the contextual objects that appeared in the most important scene views.
- SafeBench: A Safety Evaluation Framework for Multimodal Large Language Models84
SafeBench is a comprehensive framework designed for conducting safety evaluations of MLLMs that draws inspiration from the jury system in judicial proceedings and pioneer the jury deliberation evaluation protocol that adopts collaborative LLMs to evaluate whether target models exhibit specific harmful behaviors, providing a reliable and unbiased assessment of content security risks.
- 81
- Transductive Relation-Propagation Network for Few-shot Learning76
This work proposes a transductive relation-propagation graph neural network (TRPN) to explicitly model and propagate such relations across support-query pairs, the first work that explicitly takes the relations of support- query pairs into consideration in few-shot learning.
- 75
- Semantic Mirror Jailbreak: Genetic Algorithm Based Jailbreak Prompts Against Open-source LLMs70
A Semantic Mirror Jailbreak (SMJ) approach that bypasses LLMs by generating jailbreak prompts that are semantically similar to the original question is introduced and is resistant to defenses that use those metrics as thresholds.
- 66
- Harnessing Perceptual Adversarial Patches for Crowd Counting66
The Perceptual Adversarial Patch generation framework is proposed to tailor the adversarial perturbations for crowd counting scenes using the model-shared perceptual features and an adaptive crowd density weighting approach is handcrafted to capture the invariant scale perception features across various models.
- Coarse-to-Fine Image Inpainting via Region-wise Convolutions and Non-Local Correlation65
This work proposes region-wise convolutions to locally deal with the different types of regions to help exactly reconstruct existing regions and roughly infer the missing ones from existing regions at the same time, and introduces a non-local operation to globally model the correlation among different regions.
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift61
A multimodal attribution backdoor attack (MABA) is proposed that injects domain-agnostic triggers into critical areas using attributional interpretation and boosts the attack success rate of generalization over the unimodal attack.
- Manipulating Multimodal Agents via Cross-Modal Prompt Injection57
This paper proposes CrossInject, a novel attack framework in which attacker embeds adversarial perturbations across multiple modalities to align with target malicious content, allowing external instructions to hijack the agents' decision-making process and execute unauthorized tasks.
- Regionwise Generative Adversarial Image Inpainting for Large Missing Areas57
A generic inpainting framework capable of handling incomplete images with both contiguous and discontiguous large missing areas is proposed, and both qualitative and quantitative experimental results demonstrate that the proposed model significantly outperforms the state-of-the-art approaches, on the large contiguousand discontiguous missing areas.
- Pre-trained Trojan Attacks for Visual Recognition52
The Pre-trained Trojan attack is proposed, which embeds backdoors into a PVM, enabling attacks across various downstream vision tasks, and addresses the issue of shortcut connections by introducing a context-free learning pipeline for poison training.
- Unveiling the Safety of GPT-4o: An Empirical Study using Jailbreak Attacks52
This paper adopts a series of multi-modal and uni-modal jailbreak attacks on 4 commonly used benchmarks encompassing three modalities encompassing three modalities, and conducts a rigorous evaluation of GPT-4o against jailbreak attacks.
- 50
- Latent Imitator: Generating Natural Individual Discriminatory Instances for Black-Box Fairness Testing48
This paper proposes a framework named Latent Imitator (LIMI) to generate more natural individual discriminatory instances with the help of a generative adversarial network (GAN), where it imitate the decision boundary of the target model in the semantic latent space of GAN and further samples latent instances on it.
- Poisoned Forgery Face: Towards Backdoor Attacks on Face Forgery Detection47
A novel and previously unrecognized threat in face forgery detection scenarios caused by backdoor attack by embedding backdoors into models and incorporating specific trigger patterns into the input, attackers can deceive detectors into producing erroneous predictions for forged faces.
- 47
- Visual Adversarial Attack on Vision-language Models for Autonomous Driving46
This paper proposes ADvLM, the first visual adversarial attack framework specifically designed for VLMs in AD, and introduces Semantic-Invariant Induction, which uses a large language model to create a diverse prompt library of textual instructions with consistent semantic content, guided by semantic entropy.
- Compromising Embodied Agents with Contextual Backdoor Attacks46
A dual-modality activation strategy is implemented that controls both the generation and execution of program defects through textual and visual triggers that compromise key aspects of confidentiality, integrity, and availability in embodied agents.
- Generating Transferable Adversarial Examples against Vision Transformers43
An Architecture-oriented Transferable Attacking (ATA) framework is proposed to generate transferable adversarial examples by activating the uncertain attention and perturbing the sensitive embedding to activate strong attacking transferability among diverse ViTs.
- Towards Defending Multiple $$\ell _p$$-Norm Bounded Adversarial Perturbations via Gated Batch Normalization41
Gated Batch Normalization (GBN) is proposed to adversarially train a perturbation-invariant predictor for defending multiple ℓp\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy}
- GenderBias-VL: Benchmarking Gender Bias in Vision Language Models via Counterfactual Probing39
This paper introduces the GenderBias-VL benchmark, a comprehensive dataset for occupation-related gender bias evaluation, an up-to-date leaderboard on LVLM biases, and a nuanced understanding of the biases presented by these models.
- Few-shot X-ray Prohibited Item Detection: A Benchmark and Weak-feature Enhancement Network39
The Weak-feature Enhancement Network (WEN) containing two core modules, i.e. Prototype Perception (PR) and Feature Reconciliation (FR), where PR first generates a prototype library by aggregating and extracting the basis feature from critical regions around instances to generate the basis information for each category.
- Defensive Patches for Robust Recognition in the Physical World38
A defensive patch generation framework to address problems by helping mod-els better exploit these features of robust recognition by injecting class-specific identifiable patterns into a confined local patch prior, so that defensive patches could preserve more recognizable features towards specific classes, leading models for better recognition under noises.
- 38
- Towards Understanding the Safety Boundaries of DeepSeek Models: Evaluation and Findings35
Experimental results indicate that despite their strong general capabilities, DeepSeek models exhibit significant safety vulnerabilities across multiple risk dimensions, including algorithmic discrimination and sexual content.
- Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game35
This study explores the robustness of cooperative multi-agent reinforcement learning (c-MARL) against Byzantine failures and proposes a Bayesian Adversarial Robust Dec-POMDP framework, which views Byzantine adversaries as nature-dictated types, represented by a separate transition.
- Imitated Detectors35
It is revealed for the first time that black-box victim object detectors can be easily replicated without knowing the model structure and training data and a teacher-student framework named Imitated Detector is proposed to transfer the knowledge of the victim model to the imitated model.
- Towards Robust Physical-world Backdoor Attacks on Lane Detection34
BadLANE, a dynamic scene adaptation backdoor attack for LD designed to withstand changes in real-world dynamic scene factors, and a meta-learning framework to train meta-generators tailored to different environmental conditions to mitigate the effects of environmental changes.
- Towards Benchmarking and Assessing Visual Naturalness of Physical World Adversarial Attacks34
The first step to benchmark and assess visual naturalness of physical world attacks, taking autonomous driving scenario as the first attempt is taken and Dual Prior Alignment (DPA) network is introduced, which aims to embed human knowledge into model reasoning process.
- Black-Box Adversarial Attack on Vision Language Models for Autonomous Driving30
This paper proposes Cascading Adversarial Disruption (CAD), which targets low-level reasoning breakdown by generating and injecting deceptive semantics, ensuring the perturbations remain effective across the entire decision-making chain and presents Risky Scene Induction, which addresses dynamic adaptation.
- Exploring the Relationship between Architecture and Adversarially Robust Generalization30
This paper comprehensively evaluated 20 adversarially trained architectures on ImageNette and CIFAR-10 datasets towards several adversaries, and found that Vision Transformers often yield better adversarially robust generalization.
- LanEvil : Benchmarking the Robustness of Lane Detection to Environmental Illusions29
This paper systematically design 14 prevalent yet critical types of environmental illusions that cover a wide spectrum of real-world influencing factors in LD tasks and establishes the first comprehensive benchmark LanEvil for evaluating the robustness of LD against this natural corruption.
- RobustMQ: benchmarking robustness of quantized models29
This work thoroughly evaluated the robustness of quantized models against various types of noise (adversarial attacks, natural corruption, and systematic noise) on ImageNet to provide valuable insights into the robustness of quantized models in various scenarios.
- Object Detectors in the Open Environment: Challenges, Solutions, and Outlook27
This paper aims to provide a fresh, comprehensive, and systematic understanding of the challenges and solutions associated with open-environment object detectors, thus catalyzing the development of more solid applications in real-world scenarios.
- Exploring the Relationship Between Architectural Design and Adversarially Robust Generalization27
It is revealed that the higher weight sparsity contributes significantly towards the better adversarially robust generalization of Transformers, which can be often achieved by the specially-designed attention blocks, which could help to better understand the mechanism for designing robust DNNs.
- Understanding adversarial robustness via critical attacking route27
This paper draws the close connection between adversarial robustness and critical attacking route, as the route makes the most non-trivial contributions to model predictions in the adversarial setting, which could weaken the noise propagation and improve model robustness.
- Exploring Inconsistent Knowledge Distillation for Object Detection with Data Augmentation26
This work proposes inconsistent knowledge distillation (IKD), a method which outperforms state-of-the-art KD baselines on one-stage, two-stage and anchor-free object detectors (at most +1.0 mAP).
- Improving Robust Fariness via Balance Adversarial Training25
Balance Adversarial Training (BAT) is proposed, which can significantly outperform other baselines in mitigating the robust fairness problem and adjust the attack strength and difficulties of each class to generate samples near the decision boundary for easier and fairer model learning.
- GenderCARE: A Comprehensive Framework for Assessing and Reducing Gender Bias in Large Language Models23
This work constructs GenderPair, a novel pair-based benchmark designed to assess gender bias in LLMs comprehensively, and establishes pioneering criteria for gender equality benchmarks, spanning dimensions such as inclusivity, diversity, explainability, objectivity, robustness, and realisticity.
- 23
- Faire: Repairing Fairness of Neural Networks via Neuron Condition Synthesis23
This work synthesizes the condition that can reduce the effect of features relevant to the protected attributes in the DNNs, inspired by the traditional program repair method that synthesizes proper condition checking.
- 23
- Towards Transferable Attacks Against Vision-LLMs in Autonomous Driving with Typography22
This paper proposes a dataset-agnostic framework for automatically generating false answers that can mislead Vision-LLMs' reasoning, and presents a linguistic augmentation scheme that facilitates attacks at image-level and region-level reasoning, and extends it with attack patterns against multiple reasoning tasks simultaneously.
- ARShoe21
A real-time augmented reality virtual shoe try-on system for smartphones, namely ARShoe, which adopts a novel multi-branch network to realize pose estimation and segmentation simultaneously simultaneously and develops a novel stabilization method.
- VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models20
This work introduces a multimodal instruction backdoor attack, namely VL-Trojan, that facilitates image trigger learning through active reshaping of poisoned features and enhances black-box attack efficacy through an iterative character-level text trigger generation method.
- FAIRER: Fairness as Decision Rationale Alignment20
This paper investigates fairness from the perspective of decision rationale and defines the parameter parity score to characterize the fair decision process of networks by analyzing neuron influence in various subgroups and proposes gradient-guided parity alignment, which encourages gradient-weighted consistency of neurons across subgroups.
- Interpreting and Improving Adversarial Robustness with Neuron Sensitivity.19
It is demonstrated that state-of-the-art adversarial training methods improve model robustness by reducing neuron sensitivities which confirms the strong connections between adversarial robustness and neuron sensitivity as well as the effectiveness of using sensitive neurons to build robust models.
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks18
This work is the first to formalize the T2V jailbreak attack as a discrete optimization problem and proposes a joint objective-based optimization framework, called T2V-OptJail, which consists of two key optimization goals: bypassing the built-in safety filtering mechanisms to increase the attack success rate and preserving semantic consistency between the adversarial prompt and the unsafe input prompt.
- 18
- Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks18
This paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses that directly trains a defensive model by isolating the adversary's training gradient from the expected gradient, which can effectively reduce the inference computational cost.
- Improving Robust Fairness via Balance Adversarial Training18
Balance Adversarial Training (BAT) is proposed, which can significantly outperform other baselines in mitigating the robust fairness problem and adjust the attack strength and difficulties of each class to generate samples near the decision boundary for easier and fairer model learning.
- Generate Transferable Adversarial Physical Camouflages via Triplet Attention Suppression15
The Transferable Attention Attack method is proposed to generate adversarial camouflages with strong transferable attacking ability by taking advantage of visual attention mechanism, i.e., triplet attention suppression, and outperforms state-of-the-art methods.
- 15
- Few-shot Visual Learning with Contextual Memory and Fine-grained Calibration15
To the best of the knowledge, IPN is the first work that simultaneously integrates both global and local characteristics in few-shot learning, approximately imitating the human cognition mechanism.
- Adversarial Examples in the Physical World: A Survey14
This paper thoroughly examines the characteristics of PAEs within a practical workflow encompassing training, manufacturing, and re-sampling processes, and develops a comprehensive analysis and classification framework for PAEs based on their specific characteristics, covering over 100 studies on physical-world adversarial examples.
- Universal Backdoor Attacks Detection via Adaptive Adversarial Probe14
This paper proposes the Adaptive Perturbation Generation (APG) framework to detect multiple types of backdoor attacks by adaptively injecting adversarial per- turbations and introduces a gradient-guided mask generation strategy to search for the optimal regions for adversarial attacks.
- On the Guaranteed Almost Equivalence Between Imitation Learning From Observation and Demonstration13
It is proved that LfO is almost equivalent to LfD in the deterministic robot environment, and more generally even in the robot environment with bounded randomness, and suggested that it can be safely applied in practice without sacrificing the performance compared to L fD.
- Fairness Mediator: Neutralize Stereotype Associations to Mitigate Bias in Large Language Models12
This work proposes Fairness Mediator (FairMed), an effective and efficient bias mitigation framework that neutralizes stereotype associations and maintains the LLM’s language understanding capabilities without compromising overall performance.
- RoboSafe: Safeguarding Embodied Agents via Executable Safety Logic12
RoboSafe is a hybrid reasoning runtime safeguard for embodied agents through executable predicate-based safety logic that is both interpretable and executable as code, while maintaining near-original task performance.
- Bench2ADVLM: A Closed-Loop Benchmark for Vision-language Models in Autonomous Driving11
Bench2ADVLM establishes a hierarchical evaluation pipeline that seamlessly integrates high-level abstract reasoning, mid-level simulation actions, and low-level real-world execution in a closed-loop setting.
- PTSBench: A Comprehensive Post-Training Sparsity Benchmark Towards Algorithms and Models11
This work proposes the first comprehensive post-training sparsity benchmark called PTSBench, which can provide new observations for a better understanding of the PTS algorithms, in-depth and comprehensive evaluations for the sparsification ability of models, and a well-structured and easy-integrate open-source framework.
- Face Encryption via Frequency-Restricted Identity-Agnostic Attacks11
A frequency-restricted identity-agnostic (FRIA) framework to encrypt face images from unauthorized face recognition without access to personal information and outperforms other state-of-the-art methods in generating more natural encrypted faces while attaining high black-box attack success rates of 96%.
- Hierarchical Perceptual Noise Injection for Social Media Fingerprint Privacy Protection10
The proposed FingerSafe is the first to provide feasible fingerprint protection in both digital and realistic scenarios, and inject protective noises into the fingerprint orientation field to perturb the model-shared high-level semantics.
- Sequential alignment attention model for scene text recognition10
This paper proposes a sequential alignment attention model to enhance the alignment between input images and output character sequences and shows that this method substantially outperforms the state-of-the-art methods.
- 10
- ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks9
The dual-learning hypothesis is proposed, for the first time, which posits that LLMs simultaneously learn both the task-relevant latent concepts and backdoor latent concepts within poisoned demonstrations, jointly influencing the probability of model outputs.
- Module-wise Adaptive Adversarial Training for End-to-end Autonomous Driving9
This paper takes the first step in adversarial training for end-to-end AD models and presents a novel Module-wise Adaptive Adversarial Training (MA2T), which introduces Module-wise Noise Injection and Dynamic Weight Accumulation Adaptation, which incorporates accumulated weight changes to adaptively learn and adjust the loss weights of each module based on their contributions for better balance and robust training.
- Fairness via Group Contribution Matching9
This study proposes an easy but effective group contribution matching (GCM) method to match the contribution of each subgroup in the training process, and shows that this method effectively improves fairness and outperforms other methods significantly.
- 9
- TrapFlow: Controllable Website Fingerprinting Defense via Dynamic Backdoor Learning8
TrapFlow exploits the tendency of neural networks to memorize subtle patterns by injecting crafted trigger sequences into targeted website traffic, causing the attacker’s model to build incorrect associations during training, leading to widespread misclassification across unrelated websites.
- Adversarial Generation and Collaborative Evolution of Safety-Critical Scenarios for Autonomous Vehicles8
ScenGE, a framework that can generate plentiful safety-critical scenarios by reasoning novel adversarial cases and then amplifying them with complex traffic flows, is proposed, which can be applied to large model based AV systems and deployed on different simulators and observes that adversarial training on the authors' scenarios improves the model robustness.
- CopyrightShield: Enhancing Diffusion Model Security against Copyright Infringement Attacks8
This paper provides an in-depth analysis of the spatial similarity of replication in diffusion model and leverages this key characteristic to design a method for detecting poisoning data and introduces the concept of copyright feature inversion, which aids in determining copyright responsibility and expands the application scenarios of defense strategies.
- 7
- Seeing Is Not Screening: Multimodal Hidden Instruction Attacks on Agent Skill Scanners6
SkillCamo is proposed, a document-mediated multimodal instruction attack that conceals malicious instructions within images bundled with a skill while rewriting the surrounding documentation to naturally reference those images as part of the normal workflow.
- Patch Attack for Automatic Check-out6
A bias-based framework to generate class-agnostic universal adversarial patches with strong generalization ability, which exploits both the perceptual and semantic bias of models.
- 5
- 5
- 5
- 5
- SLMQuant: Benchmarking Small Language Model Quantization for Practical Deployment4
SLMQuant is introduced, the first systematic benchmark for evaluating LLM compression techniques when applied to SLMs and reveals fundamental disparities between SLMs and LLMs in quantization sensitivity, demonstrating that direct transfer of LLM-optimized techniques leads to suboptimal results.
- Towards Noise-Robust Neural Networks via Progressive Adversarial Training.4
A more powerful training method named Progressive Adversarial Training (PAT) is proposed that adds diversified adversarial noises progressively during training, and thus obtains robust model against both adversarial examples and corruptions through higher training data complexity.
- M et A dv : A Unified and Interactive Adversarial Testing Platform for Autonomous Driving3
A key feature of MetAdv is its human-in-the-loop capability: besides flexible environmental configuration for more customized evaluation, it enables real-time capture of physiological signals and behavioral feedback from drivers, offering new insights into human-machine trust under adversarial conditions.
- CodeSimpleQA: Scaling Factuality in Code Large Language Models3
This work presents CodeSimpleQA, a comprehensive bilingual benchmark designed to evaluate the factual accuracy of code LLMs in answering code-related questions, and proposes a post-training framework combining supervised fine-tuning and reinforcement learning.
- 3
- 3
- 2
- Benchmarking the Robustness of Autonomous Driving to Environmental Illusions: A Lane Perception Perspective2
This work introduces the first benchmark, LanEvil++, for evaluating the robustness of lane perception under environmental illusions, and proposes the Multimodal Illusion Defense Approach (MIDA), which uses hard examples to improve illusion resistance.
- 2
- ME: Trigger Element Combination Backdoor Attack on Copyright Infringement2
This paper raised new datasets accessible for researching in attacks like SBD, and proposed Multi-Element (ME) attack method based on SBD by increasing the number of poisonous visual-text elements per poisoned sample to enhance the ability of attacking, while importing Discrete Cosine Transform for the poisoned samples to maintain the stealthiness.
- PromptSafe: Gated Prompt Tuning for Safe Text-to-Image Generation2
This work proposes PromptSafe, a gated prompt tuning framework that combines a lightweight, text-only supervised soft embedding with an inference-time gated control network that demonstrates strong generalization to unseen harmful categories, robust transferability across diffusion model architectures, and resilience under adaptive adversarial attacks.
- CodeChemist: Test-Time Scaling for Low-Resource Code Generation via Functional Knowledge Transfer2
The extensive experiments show that CodeChemist outperforms existing test-time scaling approaches, boosting the performance of code generation for low-resource PLs without requiring any model retraining.
- 2
- 2
- ADVM'212
This workshop aims to bring together researchers from the fields of adversarial machine learning, model robustness, and explainable AI to discuss recent research and future directions for adversarial robustness of deep learning models, with a particular focus on multimedia applications, including computer vision, acoustics, etc.
- Dialogue Policies for Learning Board Games through Multimodal Communication2
This paper presents MDP policy learning for agents to learn strategic behavior–how to play board games–during multimodal dialogues, and presents a novel data collection for training natural language modules.
- Automated Program Debugging for Multiple Bugs Based on Semantic Analysis2
An algorithm combining BMPS with a semantic based debugging method, which is aimed at locating more than control flow related faults, and has significant improvement on both accuracy and efficiency.
- Towards universal X-ray security inspection: a benchmark and stereoscopic-aware oriented prohibited item detection framework1
The first oriented X-ray prohibited detection benchmark ORXray is contributed, featuring large-scale prohibited items, high-quality images, informative annotations, and real-world scenarios, and it is demonstrated that SWEAR outperforms other state-of-the-art methods by a margin of at least 2.7%, highlighting its effectiveness.
- 1
- 1
- MASteer: Multi-Agent Adaptive Steer Strategy for End-to-End LLM Trustworthiness Repair1
MASteer is proposed, the first end-to-end framework for trustworthiness repair in LLMs based on representation engineering, which demonstrates strong robustness, generalization, and practical value for scalable, efficient trustworthiness repair.
- M2G-Eval: Enhancing and Evaluating Multi-granularity Multilingual Code Generation1
M2G-Eval enables fine-grained diagnosis of code generation capabilities and highlights persistent challenges in synthesizing complex, long-form code.
- A Comprehensive Evaluation Framework for Multi-Agent Reinforcement Learning1
This work proposes an integrated framework for robust evaluation of MARL, and implements over eight adversarial attack methods—targeting policies, states/observations, actions, rewards, and environments—covering all stages of the MDP closed-loop.
- 1
- Adversarial Attacks for Embodied Agents.1
This work generates spatiotemporal perturbations to form 3D adversarial examples, which exploit the interaction history in both the temporal and spatial dimensions and adversarially perturb the physical properties of the contextual objects that appeared in the most important scene views.
- –
- –
- –
- Technical Report on the CVPR 2026@AdvML Workshop Challenge–
The CVPR 2026@AdvML Workshop Challenge on adversarial multimodal attacks against autonomous-driving VLAs is presented, providing a practical reference for future robustness evaluation and defense design in multimodal autonomous-driving systems.
- –
- –
- Ensemble-Based Uncertainty Estimation for Code Correctness Estimation–
This work proposes Ensemble Semantic Entropy (ESE), which estimates uncertainty by evaluating the consistency of samples aggregated across an ensemble of models, and proposes a cascading test-time scaling framework Cas, which maintains performance while reducing FLOPs by 64.9% compared to single-model scaling, offering a new perspective on balancing parameter and inference scaling.
- –
- Reading Between the Pixels: An Inscriptive Jailbreak Attack on Text-to-Image Models–
Etch, a black-box attack framework that decomposes the adversarial prompt into three functionally orthogonal layers: semantic camouflage, visual-spatial anchoring, and typographic encoding, is proposed, which reduces joint optimization over the full prompt space to tractable sub-problems, which are iteratively refined through a zero-order loop.
- –
- –
- –
- –
- –
- –
- SysNoise: Exploring and Benchmarking Training-Deployment System Inconsistency–
This paper introduces SysNoise, a frequently occurred but often overlooked noise in the deep learning training-deployment cycle, and builds a holistic benchmark to quantitatively measure the impact of Sys noise on 20+ models, comprehending image classification, object detection, instance segmentation and natural language processing tasks.
- –
Publication data from OpenAlex, with missing venues and authors filled in from Crossref; citation counts are the higher of OpenAlex and Semantic Scholar, last synced 2026-10-12. One-sentence summaries under some papers are written by Semantic Scholar’s model. Citation counts may be lower than on Google Scholar, which indexes more sources.
Report an error
Wrong papers, two people merged into one, or a profile that should not be here? Tell us and we will fix or hide it.